We Built the First AI-Native Quantum Software Framework: Say Hello to Agentic TensorCircuit-NG

Source: Dev.to

Quantum computing software is notoriously hard to write.
If you want to simulate a deep quantum neural network or research a new algorithm, you need to understand Hamiltonian dynamics, Hilbert spaces and be a High‑Performance Computing (HPC) expert—dealing with GPU memory limits, vectorization, JIT compilation staging times, and tensor‑network contraction paths.

For years, we’ve provided developers with the tools to do this via TensorCircuit‑NG, our next‑generation open‑source, high‑performance quantum software framework.

Today we’re changing the paradigm: TensorCircuit‑NG is now the world’s first AI‑native quantum programming platform purpose‑built for agentic quantum research and automated scientific discovery. By embedding AI skills directly into the repository, the framework now ships with a built‑in HPC engineer, a theoretical physicist, and a technical writer.

The Paradigm Shift: Agent‑Ready Architecture 🧠

Most AI coding assistants perform “line‑by‑line” translations or generate boilerplate, which doesn’t work in quantum simulation—a poorly placed for loop can increase compilation time from seconds to hours.

Instead of endless tutorials on “best practices,” we’ve embedded our framework knowledge as Agentic Skills. After cloning the latest TensorCircuit‑NG repo you’ll see a new directory structure:

.agents/skills/
├── arxiv-reproduce/
├── performance-optimize/
├── tc-rosetta/
└── tutorial-crafter/

These are strict, engineering‑bound AI workflows. Below are the four superpowers available out of the box.

1. /arxiv-reproduce: From arXiv ID to JAX‑Accelerated Code in Minutes 📄➡️💻

The gap between reading a cutting‑edge quantum machine‑learning paper on arXiv and reproducing its code is huge.
The arxiv-reproduce skill lets you provide an arXiv link; the agent will:

• Extract the physical intent (Ansatz, Hamiltonian, loss function).
• Scale down the qubit count to fit your local machine.
• Generate idiomatic, JAX‑accelerated TensorCircuit‑NG code.
• Run black formatting, pylint linting, and execute the script, saving reproduced figures into a standardized outputs/ folder.

2. /performance-optimize: Your Built‑in HPC Architect ⚡

If a quantum script takes forever to compile or crashes with an OOM error, the performance-optimize agent scans the code, identifies bottlenecks, and:

• Replaces Python loops with jax.vmap.
• Wraps deep quantum layers in jax.lax.scan to slash JIT staging time.
• Injects jax.checkpoint to trade compute for memory during back‑propagation.
• Switches to cotengra for optimal tensor‑network contraction paths.
• Runs A/B benchmarks to demonstrate speedup.

3. /tc-rosetta: End‑to‑End Cross‑Ecosystem Translation 🌍

Migrating from older, object‑oriented quantum frameworks (e.g., Qiskit, PennyLane) to a modern, differentiable, functional framework like TensorCircuit‑NG is a steep mental shift.

tc-rosetta performs end‑to‑end intent extraction: it reads a legacy script, understands the underlying mathematics, and rewrites it from scratch using pure JAX‑native paradigms. It then executes both versions and provides a benchmark report (e.g., “Execution time reduced from 300 s to 0.2 s”).

4. /tutorial-crafter: Automated High‑Quality Documentation 📝

Writing docs is the bane of every open‑source contributor. Point tutorial-crafter at any raw TensorCircuit‑NG script, and it will:

• Analyze the physical background and code.
• Generate a narrative‑driven tutorial in Markdown and HTML.
• Chunk the code logically, add LaTeX formulas for the physics theory, and highlight HPC programming tricks (e.g., “Notice how we used vmap here instead of a loop…”).

The result rivals hand‑crafted, premium tutorials.

How to Experience the Magic ✨

Because these skills are built on an open standard, getting started is zero‑friction:

1. Clone the TensorCircuit‑NG repository.
2. Open a terminal in the repo root.
3. Invoke a skill, e.g.:

/performance-optimize examples/my_slow_circuit.py

You are no longer just writing code; you are directing an autonomous digital research team.

Welcome to the era of Agentic Quantum Software Engineering. We can’t wait to see what you discover. Check out the repo, give us a star, and let the AI handle the boilerplate while you focus on the physics! 🌌